OpenAI lancia textGrain, la filigrana invisibile per i testi IA: scopri come funziona, cosa pu dimostrare e quali implicazioni ha per editori e docenti.
Rimani aggiornato con WebMasterPoint
OpenAI lancia textGrain, una filigrana invisibile pensata per riconoscere i testi generati dallIA. Come funziona e che cosa pu davvero dimostrare? Una novit con ricadute concrete per editori e insegnanti.
OpenAI announced that textGrain, an invisible watermark for ChatGPT and Codex outputs, will roll out across the European Union in the coming weeks. The move follows the 2 August 2026 deadline set by the EU AI Act, which requires providers to mark synthetic text in a machine?readable format. Unlike visible watermarks, textGrain embeds a statistical pattern hidden in the models word choices, leaving the text looking entirely natural to human readers. The system will be active by default for EU users on all plans, while API customers worldwide can enable it voluntarily. OpenAI describes the technology as a response to regulatory pressure rather than a definitive anti?plagiarism tool, and the company acknowledges that the signal can be weakened by:
- editing,
- translation,
- or heavy paraphrasing.
How OpenAI’s textGrain watermark works
The watermark works by slightly biasing the language model toward certain tokens during generation. At each step the model normally picks the next word from a probability distribution. textGrain modifies that distribution according to a secret key, creating a subtle deviation detectable by a specialized tool. Because the bias is spread across many choices, the resulting text remains readable and does not contain hidden characters or metadata. OpenAI published benchmark results showing detection accuracy of roughly 80?percent for texts of about 200 tokens and 95?percent for 400?token passages when the false?positive rate is limited to 1?percent. Performance varies with content type: mathematical explanations and short factual answers show lower rates because the vocabulary is constrained, while open?ended prose retains the signal more reliably. The watermarks resilience is also tested by human editing. Detection rate drops with synonym substitution. Replacing 10?percent of words with synonyms reduces detection from about 92?percent to 66?percent in 400?token samples, and a 25?percent substitution pushes the rate below 20?percent. Translated text and heavily paraphrased content similarly erode the signal, meaning the system is most reliable for fresh, unmodified outputs. Watermark cannot identify user or account, nor can it trace the original prompt or conversation.
What the detection canand cannotprove
When the detector flags a passage as watermarked, it indicates that an OpenAI model likely generated the text, but it does not quantify how much of the content originated from the AI versus human editing. The signal cannot reveal:
- the users identity,
- the prompt that triggered generation,
- or whether the text was later altered by a human hand.
Consequently, a positive detection does not constitute proof of plagiarism or misuse. Conversely, a negative detection does not guarantee human authorship; short texts, code snippets, or heavily edited passages may escape the detector even when produced by an AI model. These limitations mean that textGrain should be treated as an indicator of provenance rather than a definitive verdict on authenticity.
Practical implications for publishers and educators
Publishers cannot currently offer the detector to the general public: access is limited to approved researchers and expert organizations. They can use OpenAIs API-based opt-in feature or wait for broader access before incorporating detection into their workflows.
Educators and publishers should treat a positive result as one piece of evidence, not as proof of plagiarism or a measure of how much a person contributed. Pair it with plagiarism checks, stylistic analysis and editorial judgment, and avoid drawing firm conclusions from short or substantially edited passages. The reported 1 percent false-positive threshold also means a flag can be wrong.